VLA Gallery / All models / New in 2026

ω-0 (OMEGA-0)

NTU(MARS Lab)·Peking University·BAAI·HKUST(GZ) · 2026-08

VLANew in 2026About 4Bdiffusion

A humanoid latent-predictive world action model that predicts future observation embeddings from language·egocentric/third-person video·proprioception while generating SONIC-compatible whole-body action latents with an action DiT.

Paper

ω-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-ManipulationarXiv 2608.06375

Overview

Category

New in 2026

Type

VLA

Released

2026-08

Organization

NTU(MARS Lab)·Peking University·BAAI·HKUST(GZ)

Open weights

Yes

Parameters

About 4B (HF metadata; VLM Qwen3-VL-2B, action DiT 0.45B (paper Fig. 2), total not stated in paper)

Architecture

Backbone LLM

Qwen3-VL-2B-Instruct (whole-body action VLM; LLM not separately stated)

Vision encoder

V-JEPA 2.1 image encoder (frozen) + Qwen3-VL built-in vision

Action head

Action DiT diffusion (x0 prediction, few DDIM steps) → 66-dim latent (SONIC 64 + hands 2) (diffusion)

Chunk

25 steps (8 executed) — public config 30

Control rate

>7 Hz (forward pass about 0.14 s)

Architecture (original summary)

Qwen3-VL-2B whole-body action VLM (FAST token pretraining) + V-JEPA 2.1 (frozen)·T5 → joint video-action latent predictor (future Wan latent L2) → action DiT diffusion (x0 prediction, DDIM) → 66-dim (SONIC whole-body latent 64+hands 2) chunk H=25, 8 executed, >7 Hz → SONIC controller

Training

Pretraining accelerators

8 H100 (stages 1·2; stage 3 same)

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Stage 3 real-data fine-tuning (11 tasks about 2,220 trajectories, VLM · V-JEPA · Wan frozen) on 8 H100, time and steps not disclosed (public config defaults: batch 32 · 80k steps · lr 1e-4)

Data type

Real-robot teleoperation / human video / simulation·synthetic

Data scale

ω-HOME 40.3 hours·4,827 episodes·24 tasks (30 Hz); downstream 11 tasks about 2,220 trajectories (about 200 per task); scale of public human data (ARCTIC·Xperience-10M·Motion-X) not stated

Main datasets

ω-HOME; ARCTIC; Xperience-10M; Motion-X

Tasks & evaluation

11 real-world household loco-manipulation tasks (10 trials per task): success rate 81.8% (Omni)/79.1% (Ego), score 36.7/41, progress 90.3%; baselines ψ-0 44.5%, DiT4DiT 43.6%, Fast-WAM 37.1%, InternVLA-M1 31.8%, π0.5 27.3%, GR00T-N1.7 22.7%

Robot

Unitree G1 humanoid (README; paper says 'G1') + Inspire DexHand, head ZED Mini egocentric camera; external ZED depth camera during collection

Inference

Inference device & speed

About 0.14 s per forward pass (>7 Hz), K=8 of chunk H=25 executed + RTC-style warm start · overlap blending; inference server GPU model not stated, SONIC on workstation or G1 Orin (README)

Inference memory

Not disclosed

Weights bf16 (calc.)

8 GB

Where it runs

Workstation GPU inference server (model not stated) + SONIC (workstation or G1 Orin)

Reinforcement learning & base

Reinforcement learning

Partial (RL low-level controller) · ω-0 itself is imitation learning (FAST token NTP + diffusion x0 prediction); SONIC tracking controller trained with PPO executes the latents, and training action latents are generated by SONIC sim replay of public motions

RL evidence

https://arxiv.org/abs/2511.07820

Base control

Humanoid legged (Commands/latents → RL locomotion controller) · VLA outputs chunks of SONIC-compatible whole-body action latents (64-dim) + hand commands → SONIC RL tracking controller handles locomotion·whole-body joint control

Notes

Key features

A humanoid latent-predictive world action model that predicts future observation embeddings from language·egocentric/third-person video·proprioception while generating SONIC-compatible whole-body action latents with an action DiT. On 11 real-world household loco-manipulation tasks it reached 81.8% success, far ahead of baselines such as ψ-0 (44.5%)·π0.5 (27.3%).

Notes

Public human motions converted into robot-executable action latents·states via SONIC sim replay, with tracking-failure trajectories removed; teleoperation drives SONIC with Pico 4 Ultra·Pico trackers; chunk 25 (paper) vs 30 in public config; code (GitHub MIT, HF card Apache-2.0)·weights (keycharon/Omega-zero)·ω-HOME (HF 38.2GB) released; arXiv v2 2026-08-09. Based on: SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control (arXiv 2511.07820) — executes via SONIC action latent space·controller and latentizes public motions via SONIC replay (stated in paper)

Verification

Corrections applied